USEFUL: mUltimodal Sensing and pErception For autonomoUs vehicLes

USEFUL is a multimodal autonomous driving perception dataset featuring 10 synchronized sensor streams — LiDAR, 2x Radar, 3x RGB cameras, Thermal (LWIR), SWIR, and Polarimetric — paired with accurate 3D and 2D bounding box annotations and GPS/INS ego-pose data. It is designed to advance research in challenging perception conditions (night, glare, fog, rain) where passive imaging alone is insufficient.

---

Key Statistics

Total size	~345 GB
Scenes	83
Samples	10,953
Sensor streams	9 synchronized channels
3D annotations	86,662
2D annotations	636203
Instances	6274
Object categories	10
Splits	train / val / test

---

Sensor Modalities

Channel	Modality	Description
LIDAR	LiDAR	3D point cloud, XYZIRGB format
RADAR_LEFT	Radar	Left-side radar with Doppler velocity
RADAR_RIGHT	Radar	Right-side radar with Doppler velocity
WIDE_LEFT	RGB Camera	Wide-angle left camera
NARROW	RGB Camera	High-resolution narrow front camera
WIDE_RIGHT	RGB Camera	Wide-angle right camera
LWIR	Thermal	Long-wave infrared (thermal) camera
SWIR	SWIR	Shortwave infrared camera
POLARIMETRIC	Polarimetric	Full Stokes polarimetric camera (DOLP, AOLP, RGB 0°/45°/90°/135°)

The LiDAR sensor used is not a conventional rotatory LiDAR. It is an L3CAM, produced by Beamagine S.L., multimodal embbeded system containing a thermal camera, a polarimetric camera and a MEMs-based quasi-solid state LiDAR that produce high-density point clouds within a FOV of (60º, 20º).

---

Object Categories

pedestrian · vehicle.car · vehicle.motorcycle · vehicle.construction · vehicle.bus · vehicle.truck · sign · bicycle · personal_mobility · animal

---

Data Structure

Repository layout

GerardDMG/USEFUL/
├── shards/          # Raw sensor data — WebDataset TAR archives (.tar)
└── v1.0/            # Metadata — 12 JSON tables

• shards/ — WebDataset TAR archives containing the raw sensor files (point clouds and images) for each sample.
• v1.0/ — Twelve JSON files, one per metadata table. These do not require decompression.

Metadata tables (v1.0/)

Every record is uniquely identified by a 32-character hex token. Tables are linked to each other via these tokens.

Table	Description
log	Top-level recording sessions (vehicle, date, weather, location)
scene	Short clips within a log; linked list of samples
sample	Synchronized multimodal frame; linked list via next/prev
sample_data	One sensor file per channel per sample
sample_annotation	3D bounding box in the LiDAR/ego frame
sample_annotation_2d	2D bounding box in the camera frame
sensor	Sensor metadata (modality + channel name)
calibrated_sensor	Extrinsic and intrinsic calibration per sensor
instance	Object identity tracked across frames
category	Object class definitions
visibility	Per-annotation visibility level (1–4)
ego_pose	Vehicle GPS/INS/IMU pose at each sample timestamp

Sensor file naming convention

{vehicle}-{datetime}__{CHANNEL}__{timestamp_us}.{ext}

Example:

u001-2024-11-19-12-46-25+0100__LIDAR__1732020526668479.pcd
u001-2024-11-19-12-46-25+0100__NARROW__1732020526668479.png

File formats: .pcd for LiDAR/Radar point clouds, .png for all camera modalities except for POLARIMETRIC, which is .bmp.

---

Download

Install the HuggingFace Hub library and use snapshot_download to fetch the dataset:

from huggingface_hub import snapshot_download

# Download the full dataset (~345 GB)
snapshot_download(
    repo_id="GerardDMG/USEFUL",
    repo_type="dataset",
    local_dir="./USEFUL/shards/",
    local_dir_use_symlinks=False
)

Selective download (skip sensors you don't need)

Use ignore_patterns to skip specific sensor modalities and save disk space:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="GerardDMG/USEFUL",
    repo_type="dataset",
    ignore_patterns=["*SWIR*", "*WIDE_LEFT*", "*NARROW*"],  # skip these sensors
    local_dir="./USEFUL/shards/",
    local_dir_use_symlinks=False
)

Valid pattern keywords correspond to sensor channel names: LIDAR, RADAR_LEFT, RADAR_RIGHT, WIDE_LEFT, NARROW, WIDE_RIGHT, LWIR, SWIR, POLARIMETRIC.

---

Decompression

Each sensor's data is distributed as split multi-part archives with optional MD5 checksums for integrity verification:

shards/
├── LIDAR.tar.aa
├── LIDAR.tar.ab
├── LIDAR.tar.md5
├── NARROW.tar.aa
├── NARROW.tar.ab
├── NARROW.tar.md5
└── ...

Use the provided decompress_useful.py script to reconstruct and extract the archives. It automatically verifies MD5 integrity (if checksum files are present) before extracting.

# Extract all sensors
python decompress_useful.py

# Extract specific sensors only
python decompress_useful.py --sensors LIDAR NARROW LWIR

By default, files are extracted to a samples/ directory. The script handles both plain (.tar.*) and gzip-compressed (.tar.gz.*) split archives.

The metadata in v1.0/ consists of plain JSON files and requires no decompression.

After extraction, point the devkit at the directory containing v1.0/ as the dataroot.

---

Useful-devkit

The USEFUL devkit provides a Python API to load, query, visualize, and export all dataset content.

GitHub: https://github.com/GDMG99/useful-devkit

git clone https://github.com/GDMG99/useful-devkit.git
cd useful-devkit
pip install -e .

---

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Non-commercial use only.

---

Acknowledgements

This work would not have been possible without the open-source works of nuScenes and Truckscenes.

---

Citation

If you use the USEFUL dataset in your research, please cite using our DOI.